llms.txt for Manufacturers
What the llms.txt convention proposes, which engines are documented to read it as of August 2026, and why a generated file is worth publishing while a hand-maintained one is worse than none.
llms.txt is a proposed convention placing a markdown index of a site's key pages at the domain root. No major engine documents consuming it as of August 2026. A generated one costs nothing and a hand-maintained one drifts, which makes generation the only version worth publishing.
Two files sit at the root of a site and both get discussed as AI controls. They do different things. llms.txt is an offer, and robots.txt is a permission. Only one of them is read by the engines, and it is not this one.
What the convention proposes
llms.txt is a markdown file at the domain root listing a site's key pages with a one-line description of each. It was proposed in September 2024 by Jeremy Howard of Answer.AI, on the reasoning that a model working within a context window benefits from a curated map more than from a crawl.
The format is deliberately plain: an H1 with the site name, a blockquote summary, then sections of markdown links with descriptions. A companion llms-full.txt carries full page text rather than links.
The idea is coherent. A supplier site with 400 part pages and six capability pages does have a small subset worth reading first, and no existing file says which.
What is documented, as of this date
No major engine publishes documentation stating that it fetches or uses llms.txt. That is the state on 2026-08-08 across OpenAI, Anthropic, Google, and Perplexity, and it is the fact that should decide how much effort the file gets.
Google search relations staff have publicly likened it to the keywords meta tag, a signal sites declared for years and search engines never consumed. That is a comparison rather than a policy statement, and it is the closest thing to an official position anyone has offered.
This is the same shape as any absence, and it carries the same caveat the measurement work does: no documentation is not proof of no consumption. Engines do not publish their full retrieval behaviour. What can be said is that nobody has committed to reading it, and that a convention with no committed consumer should be costed accordingly.
If an engine documents support, this page is wrong and should be updated rather than defended.
What is actually fetched
The page. A browsing agent following a link or a search result retrieves the URL, and what it gets is the HTML.
That makes the adjacent practice more useful than the file itself. A markdown mirror serves a clean equivalent of each page at the same URL plus .md, so an agent that prefers markdown takes the text without parsing navigation, cookie notices, and layout. This site publishes one for every indexable URL, and the same reasoning makes a technical spec sheet PDF worth pairing with an indexable HTML equivalent.
That is a retrieval improvement with a mechanism behind it. llms.txt is a directory of things to retrieve, which only helps something that reads directories.
Why a hand-maintained file is worse than none
The failure mode is drift, and it is not hypothetical.
A file naming URLs that redirect advertises 308s to the only audience that reads it. This site's own llms.txt did exactly that before it was rebuilt: it walked the content directory independently, named zero of the eight Core pages, and listed two URLs that answered with redirects.
The fix was structural rather than editorial. llms.txt on this site now derives from the same module that feeds the sitemap and the routes, so a page cannot appear in one and be missing from another. Three consumers, one source.
The property that makes drift possible is worth naming, because it is not specific to this convention. File drift follows from a second inventory, and any root-level file listing pages is a second inventory unless something generates it.
A manufacturer publishing llms.txt from a CMS plugin that generates it has that property. One maintaining a static file by hand acquires a second inventory of the site that starts diverging on the next page launch, and it will diverge silently, because nothing checks it.
llms.txt grants no access
Listing a page in llms.txt does not make it accessible to a crawler blocked in robots.txt. The file expresses a preference about what to read, not a permission to read it, and the two are enforced by nothing and by convention respectively.
That distinction is the reason these are separate decisions. A manufacturer that blocks retrieval agents at the edge and publishes a careful llms.txt has produced a reading list for visitors who cannot get in. The access decision is covered in AI crawler access and it is the one with consequences.
What to do
Publish it if it is generated. The cost is a build step and the downside is bounded.
Do not hand-maintain it. A stale index is a worse artifact than no index, and no one will notice it went stale.
Do not count it as AI visibility work. Generative engines cite sources carrying explicit, extractable, attributed statements, and no root-level file substitutes for having those statements on the page. The composition that does the work is in citation-worthy content structure, and the selection it feeds is in how LLMs choose suppliers.
Spend the equivalent effort on schema and on specifications in text. Schema markup has documented consumers where llms.txt does not: Organization, Product, and FAQ types are parsed by search systems that publish what they read and what they do with it. That is the difference worth acting on, and the implementation belongs with the rest of the technical work.
Common questions
Is llms.txt a standard?
No. It is a proposal with an author and a specification, adopted voluntarily. Nothing ratifies it and no engine has committed to it.
Does publishing one hurt?
Only through drift. A generated file is harmless. A file that has been wrong for a year is a small, quiet misrepresentation of the site to the one audience that would take it literally.
Should llms-full.txt be published too?
Rarely worth it for a manufacturer. It duplicates the site as a single document, and where a page is already served as markdown the duplication buys nothing.
Does it help with AI Overviews?
No mechanism connects them. Overviews are assembled from the search index, which is built by ordinary crawling.
If nobody reads it, why does this site publish one?
Because it is generated from an existing source at no marginal cost, and because a convention with a plausible mechanism and no documented consumer is exactly the kind of thing worth having in place cheaply rather than adopting expensively later.
Sources
- 01
Last reviewed . Published by ManufacturingSEO.ai.